1 / 26

Personalize Expedia Hotel Searches –2 nd Place

Personalize Expedia Hotel Searches –2 nd Place. Jun Wang Alexandros Kalousis. Outline. Problem Setting Modelling Feature Engineering Conclusion and Future Work. Learning Problem.

mignon
Download Presentation

Personalize Expedia Hotel Searches –2 nd Place

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Personalize Expedia Hotel Searches –2nd Place Jun Wang Alexandros Kalousis

  2. Outline • Problem Setting • Modelling • Feature Engineering • Conclusion and Future Work

  3. Learning Problem • Learning to rank hotels such that purchased and clicked hotels are ranked at top among all hotels associated with a search query.

  4. Ranking Model

  5. Assumption • The relevance score reflects user’s rational behaviour based on the quality of hotels provided by Expedia. • Any kind of irrational behaviour of user is treated as random noise in the data

  6. Outline • Problem Setting • Modelling • Feature Engineering • Conclusion and Future Work

  7. Modelling Methodology • Linear – difficult to handle categorical (discrete) features • Linear Regression • linear learning-to-rank model (RankSVM and SGD approach) • Nonlinear • Gradient Boosted Trees • RandomForest • LambdaMART

  8. LambdaMART • LambdaMART is a learning to rank algorithm based on Multiple Additive Regression Tree (MART). • It is nonlinear and computationally efficient. NDCG score of , with j, k elements swapped NDCG score of This objective is minimized by learning relevance score through MART.

  9. Outline • Problem Setting • Modelling • Feature Engineering • Conclusion and Future Work

  10. Feature Engineering (1) • What is a good representation forLambdaMART(tree-based classifier) ? Y< b X< a Feature Y is more discriminative than feature X for tree classifier

  11. Feature Engineering (2) • We want feature with monotonic utility w.r.t. target variable!

  12. Feature Engineering Tasks • Missing value estimation • On hotel descriptions • On user’s historical data • On competitor descriptions • Feature extraction • Feature normalization

  13. Missing Value Estimation (1) Hotel Descriptions • User do not like to book and click hotels with missing values. • Our solution: fill missing value of hotel description with worse case scenario. booking_bool click_bool

  14. Missing Value Estimation (2) User’s Historical Data • Approximately 95% of user’s historical data are missing!!! • No enough information to model user’s historical data • Our solution: highlight the matchingand (or) mismatchingbetween historical data and the given hotel data

  15. Missing Value Estimation (3) Competitor Descriptions • Missing value of competitor descriptions are all set to zero • In total, there is no significant hotel price difference between Expedia and other competitors.

  16. Feature Engineering Tasks • Missing value estimation • Feature extraction • Hotel quality estimation • Non-Monotonicity of Feature Utility • Feature normalization

  17. Feature Extraction (1) • On user’s historical data • On hotel quality • The probability of each hotel being booked or clicked varies very much. • This probability is estimated by and The number of times that prop_id was booked The number of times that prop_id appeared in the data

  18. Feature Extraction (2) • On the Non-Monotonicity of Feature Utility • Some features have non-monotonicity utility functions. • E.g. prop_starrating

  19. Feature Engineering Tasks • Missing value estimation • Feature extraction • Feature normalization

  20. Feature Normalization (1) • Market, e.g. hotel price, varies in different cities, at different times. • To build a good ranking model for all hotels worldwide, corresponding features should be normalized in an appropriate manner. • Removing scaling factors

  21. Feature Normalization (2) • Our solution: normalize hotel and competitor descriptions with respect to different indicators • srch_id: compare the quality of hotels in the impression list • prop_id: compare the quality of hotels in their own history • srch_des_id: compare the quality of hotels in a given region • month: compare the quality of hotels in a given time • srch_book_window: compare the quality of hotels for a given booking window • …

  22. Example -- Hotel Price srch_id prop_id srch_destination_id

  23. Learning • Training Data • Num. of instance: 9M • Training data 80% and Validation data 20% • Num. of feature: 300 (after feature engineering) • Methodology • Linear model: regression or SVM-Rank • Nonlinear model: LambdaMART

  24. NDCG Roadmap LambdaMART + feature normalization w.r.t. a number of indicators LambdaMART + feature extraction LambdaMART + feature normalization w.r.t. srch_id Linear Model: Regression or SVM-Rank

  25. What matters? • Feature Engineering • Linear vs. Nonlinear • Number of learning instances

  26. Future works • User’s behaviour is not rational • Modelling the position bias • Feature engineering • Missing value estimation in a more principle way • Normalizing features with multiple indicators • Remove redundant features • Modelling Methodology • Improving LambdaMART • New modelling formulation based on MART

More Related